Multi-feature Fusion For Image Retrieval

With the popularity of the Internet and the rise of mobile smart terminals,more and more image video data was created in recent decades.Facing the massive multimedia data,people are no longer simply satisfied with the image retrieval through specific text.More people want to get their own pictures of interest with image retrieval through a specific picture.In content-based image retrieval,feature representation is a fundamen-tal issue and the selection of features significantly impacts the retrieval performance.Generally,the performance of a single visual feature is limited and the fusion with mul-tiple complementary features is a preferable solution to boost the accuracy of image retrieval.In order to fuse the search results based on different features of the image,we have two key problems to be solved.The first key problem is how to make the distance metric based on different feature spaces comparable.Because the distance calculated on different features such as SIFT,HSV,CNN is usually not in a scale space.It is inappropriate to add the distance directly that is not in the same scale space.The second key problem is how to adaptively evaluate the effectiveness of different features and quantify them digitally.Because for some of query images,the local feature could achieve better search results.However,for some other query images,the global feature such as CNN feature may obtain better retrieval results.For those features that have a better retrieval performance,a greater weight should be given to them when we search the relevant images though several features.Based on the above two key problems,our work is summarized as follows:(1)Adaptively weighted graph fusion for image retrieval.In this method,the dis-tances between different feature spaces are unified into a graph,in which the similarity between the images is measured by Jaccard coefficient.At the same time,in order to measure the effectiveness of different features,we use the PageRank algorithm to ana-lyze the graphs of different features,and measure the effectiveness of different features by analyze the distribution of the finally obtained PageRank values.Finally,the adap-tive weighting fusion based on the graph model is completed.According to the final fused graph,we can get the rank list for the image retrieval.(2)Adaptive multi-feature fusion based on neighborhood similarity distribution for image retrieval.This method for feature fusion is based on the distribution of the neighboring space of the images.The k neighborhood's spatial distribution is not the same for a specific query image with different features.We explore the spatial distribu-tion to compute the effectiveness of different features.We propose Rank Effectiveness Coefficient(REC)to evaluate the effectiveness of the similarity scores for a specific im-age and the feature.The similarity scores of the image with different features is fused based on the Rank Effectiveness Coefficient.According to the fused similarity scores we can obtain the final sort results for the query image.